Beta hebbian learningdefinition and analysis of a new family of learning rules for exploratory projection pursuit

  1. Quintián Pardo, Héctor
Dirigida por:
  1. Emilio Santiago Corchado Rodríguez Director/a

Universidad de defensa: Universidad de Salamanca

Fecha de defensa: 09 de junio de 2017

Tribunal:
  1. Pablo García Bringas Presidente
  2. Leticia Elena Curiel Herrera Secretario/a
  3. Ajith Abraham Vocal

Tipo: Tesis

Teseo: 483354 DIALNET

Resumen

This thesis comprises an investigation into the derivation of learning rules in artificial neural networks from probabilistic criteria. •Beta Hebbian Learning (BHL). First of all, it is derived a new family of learning rules which are based on maximising the likelihood of the residual from a negative feedback network when such residual is deemed to come from the Beta Distribution, obtaining an algorithm called Beta Hebbian Learning, which outperforms current neural algorithms in Exploratory Projection Pursuit. • Beta-Scale Invariant Map (Beta-SIM). Secondly, Beta Hebbian Learning is applied to a well-known Topology Preserving Map algorithm called Scale Invariant Map (SIM) to design a new of its version called Beta-Scale Invariant Map (Beta-SIM). It is developed to facilitate the clustering and visualization of the internal structure of high dimensional complex datasets effectively and efficiently, specially those characterized by having internal radial distribution. The Beta-SIM behaviour is thoroughly analysed comparing its results, in terms performance quality measures with other well-known topology preserving models. • Weighted Voting Superposition Beta-Scale Invariant Map (WeVoS-Beta-SIM). Finally, the use of ensembles such as the Weighted Voting Superposition (WeVoS) is tested over the previous novel Beta-SIM algorithm, in order to improve its stability and to generate accurate topology maps when using complex datasets. Therefore, the WeVoS-Beta-Scale Invariant Map (WeVoS-Beta-SIM), is presented, analysed and compared with other well-known topology preserving models. All algorithms have been successfully tested using different artificial datasets to corroborate their properties and also with high-complex real datasets.